Harness is a meta-skill for Claude Code that automates the design and generation of domain-specific agent teams and their associated skills. It addresses the complexity of coordinating multiple specialized AI agents for intricate tasks, enabling users to quickly set up sophisticated workflows. The primary benefit is the significant improvement in LLM code agent output quality and reliability through structured pre-configuration.

How It Works

Harness leverages Claude Code's agent team system to decompose complex problems into coordinated, specialized agents. It automatically generates agent definitions (in .claude/agents/) and skills (in .claude/skills/) tailored to a specific domain. The system supports six distinct agent team architectural patterns: Pipeline, Fan-out/Fan-in, Expert Pool, Producer-Reviewer, Supervisor, and Hierarchical Delegation. Skills are auto-generated with Progressive Disclosure for efficient context management, and Harness includes orchestration protocols for inter-agent data passing, error handling, and team coordination, along with validation and testing mechanisms.

Quick Start & Requirements

• Installation:
  • Via Marketplace: /plugin marketplace add revfactory/harness followed by /plugin install harness@harness.
  • Direct: Copy the skills/harness directory to ~/.claude/skills/harness.
• Prerequisites: Requires Claude Code with experimental agent teams enabled: CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1.
• Links: No external quick-start or documentation links are provided beyond the README.

Highlighted Details

• Implements 6 core agent team architectural patterns for flexible task decomposition.
• Automates the generation of agent definitions and specialized skills using Progressive Disclosure.
• Includes built-in validation, dry-run testing, and comparison tests for generated agent workflows.
• The harness-100 project offers 100 pre-built agent team harnesses across 10 domains.
• A/B testing demonstrates a +60% improvement in average quality score and a 100% win rate for LLM code agent tasks compared to un-harnessed approaches, with effectiveness scaling with task complexity.

Maintenance & Community

No specific details regarding maintainers, community channels (like Discord/Slack), or roadmaps are provided in the README.

Licensing & Compatibility

• License: Apache 2.0.
• Compatibility: The Apache 2.0 license is generally permissive for commercial use and integration into closed-source projects.

Limitations & Caveats

The system requires enabling experimental features within Claude Code (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1), suggesting it may be in an early or beta stage of development and subject to change.